Open by Strategy

The United States built the best AI models and mostly locked them behind paid interfaces. China took the opposite path: it gave its models away, releasing the actual weights for anyone to download, run and modify. That choice has made Alibaba’s Qwen the most downloaded open AI model family in the world.

The scale of adoption is remarkable. On the developer platform Hugging Face, Qwen models have been downloaded more than two billion times and have spawned over 150,000 derivative models, roughly two and a half times as many as Meta’s Llama and nearly double Google’s, from a family of more than 460 released models.^[1][2] When DeepSeek released its R1 reasoning model in January 2025 under a permissive open licence, performance near OpenAI’s best, at a fraction of the price, it triggered a $600 billion single-day fall in Nvidia’s value, the largest in US market history.^[3][4]

Giving the model away is not generosity. It is how you make hundreds of thousands of developers build on your platform for free.

The platform play

This is strategy, not charity. Chinese labs, Alibaba’s Qwen, DeepSeek, Moonshot’s Kimi, Zhipu’s GLM, ship open weights precisely because it wins the ecosystem: every developer who fine-tunes an open model, every derivative built on it, deepens dependence on that family and pulls workloads toward the cloud that hosts it best. Against closed American models, open weights are how a challenger builds a platform — by making the barrier to adoption zero.

One honest caveat on the numbers. Download counts are not strictly comparable across platforms, Alibaba’s own tally mixes Hugging Face with its ModelScope hub, and downloads are not the same as active usage. The Hugging Face-verified figures are the rigorous ones; the “most-downloaded” claim survives them.

What to watch from here

Open-weight models are China’s most effective platform strategy in AI, and their significance runs well beyond model quality. By giving models away, Chinese labs are buying something more valuable than licence fees: developer mindshare, a global base of fine-tuned derivatives, and a gravity well that pulls AI workloads toward Chinese clouds and tooling. For investors that reframes the AI-platform contest; the question is not only whose model is smartest but whose ecosystem developers actually build on, and on that measure China’s open strategy is winning share the closed American approach cannot easily contest. We treat Qwen’s download lead as an ecosystem asset, not a vanity metric, the foundation of a platform, given away to be owned.